Management Journal for Advanced Research

1. Introduction

According to International Diabetes Federation [1], type 2 diabetes (T2D) is a worldwide epidemic affecting 8.3% of the global population and the total number of people living with diabetes is projected to rise to643 million people by 2030and783 million people by 2045. To make it worse, almost half of people with diabetes are unaware that they are living with the condition and three in four adults with diabetes live in low- and middle-income countries. In U.S. alone, 38.4 million people (or 11.6% of the U.S. population) suffer from diabetes and its complications, and, of those 38.4 million U.S. individuals, 8.7 million (22.8%) are undiagnosed [2].

Note that diabetes can cause serious complications such as hyperglycemia, heart disease, blindness and even death if it is not treated in time and properly. However, it is very difficult to treat T2D in time because T2D-related complications often develop before T2D is diagnosed by up to 10 years [3]. Moreover, diabetes is a high-cost epidemic and hence individuals with diabetes spend an average annual medical cost of $19,735, which is 2.6 times higher than what individuals without diabetes pay.According to the Economic Report from the American Diabetes Association in 2022, the total annual cost of diabetes in U.S. for 2022 is $412.9 billion, including $306.6 billion in direct medical costs and $106.3 billion in indirect costs (i.e., disability, work loss or premature death).

It is also important to note that T2D is a chronic medical condition with a high risk of readmission even after proper treatment. Because of its difficulty to early detect and completely treat, the number of people with T2D has been increasing and it is projected that one of every three Americans will have diabetes in 2050.

Therefore, it will be of great value to identify biological, genetic, cultural, behavioral and demographic factors that increase the risk of developing T2D. Once such factors are identified, it is very possible to develop predictive models to pinpoint who are most likely to develop T2D even when they are very young. Then the next step is to develop a comprehensive T2D intervention program to carefully monitor T2D patients and people with a high risk of developing T2D through diabetes specialist teams during their hospitalization and after discharge.

To this end, several studies employed statistical models to identify possible causal relationships between variables and T2D incidents from small data sets. For example, Wilson et al. [7] provided the prediction rules of T2D incidents for middle-aged white adults using a series of logistic regression models. They calibrated two models so that their simple statistical model included only limited personal information such as age, gender and parental history of diabetes, while their more complex model included additional information from clinical experiments such as measurement of insulin levels. Through their comparative experiments, they concluded that several variables including parental history of diabetes, obesity, elevated triglyceride levels, low levels of high-density lipoprotein cholesterol, and hypertension statistically are significant predictors of T2D incidence. In a study from [8], separate regression models were applied to each gender group to identify informative variables from a set of clinical, biological, and genetic information obtained from patients in 30–65 years of age. They found that several clinical (e.g., adiposity) and biological variables (e.g., baseline glucose) were statistically significant while genetic variables were only marginally predictive.

Other studies have utilized the rapid advancement of genotyping technologies to investigate predictors of incident diabetes. For example, a classical study showed that T2D is strongly heritable with estimates ranging from 30% to 70% based on their observations of concordance for T2D diabetes mellitus in male twins [9]. In addition, Saxena et al. [10] identified three additional diabetes-associated loci in multi-ethnic populations based on a large meta-analysis of T2D-candidate-gene association studies. In a study [11], 70 diabetes-associated loci have been identified although signals originated from such genes explained only 10% of the genetic variance of T2D risks. Similarly, another study was undertaken to locate T2D loci through a gene-centric meta-analysis [12]. Another popular direction of studies is to employ clinically proved evidence to reduce the prevalence of T2D patients and lower their readmission rates. For example, Strack et. al. [13] reported that measuring hyperglycemia, a clinically effective treatment of diabetes, of diabetes patients during the hospitalization can lower readmission rate.

Another popular direction of diabetes-related studies involves in multi-ethnic populations, environments, and socioeconomic position.

biologists and geneticists, and hence did not provide financial/economic benefits and managerial insights that state and federal government officials needed to develop and initiate a comprehensive T2D intervention program. In particular, it is strongly recommended to develop data-driven prediction models from a large data collected and managed through data warehouse technology.

3. Data Sets & Prediction Tasks

The diabetes dataset used in this study consists of 101,766 inpatient encounter records along with demographic (e.g., race, gender, age, and weight) and clinical/medical test features (e.g., primary diagnosis, admission type, number of lab procedures, number of medications and number of emergency visits, and others). This data is a preprocessed version of the Health Facts data collected from 130 US hospitals and integrated delivery networks over 10 years for a prior study [13], which is available at UCI Machine Learning Repository [22]. For our further analysis, we removed three attributes—weight (97%), payer_code (52%), and medical_specialty (53%) from our analysis due to high rate of missing values (>50%), resulting in 14 nominal and eight numerical input variables in addition to the class variable, “readmitted.”

The classification task in this study is to predict one of three possible values of readmitted variable for each record: “<30” for a readmitted patient within 30 days after being discharged from a hospital, “>30” for a readmitted patient after 30 days from discharge, and “No” for a patient who is not readmitted at all. Three most representative data mining models (artificial neural network (ANN), decision tree (DT), and Naïve Bayes (NB)) were built with the default parameter setting of Analysis Server in Microsoft SQL Server 2014. We randomly select 70% (21,000 records) of records for calibrating each model and the remaining 30% (9,000 records) for testing purpose. In the final data set, patients with “<30”, “>30”, and “No” class variable values consist of 11.56%, 32.97% and 55.47% of records, respectively.

Note that, in this study, it is not our main goal to develop a completely new prediction algorithm or fine tune any existing algorithm to obtain a better performing algorithm.

By definition, a lift chart for a random model is always a diagonal line because it will statistically correctly identify x% of positive records when top x% of population are classified.

4. Model Selection & Analysis

4.1 Model Selection

Based on values of four numerical metrics, we immediately removed NB from further consideration mainly because its low accuracy and specificity value. We also consider the fact that while many data sets contain numerical variables but NB cannot process them appropriately by its nature. While ANN and DT are comparable in terms of accuracy (68.77% vs. 69.35%), we preferred ANN due to its significantly higher sensitivity value (10.00% vs. 3.98%). Note that sensitivity indicates a model’s ability to correctly detect patients who are readmitted (or, in general, who have a disease) and is preferred to specificity in a medical domain. Therefore, from now on, visual outcomes from ANN only will be presented due to the limited space. However, our analysis structure is generic and hence outputs from DT can be presented and discussed in the same manner.

Using ANN as a prediction model, three lift charts can be obtained depending on the prediction task. For example, the first and the second lift chart in Figure 1-a and 1-b represents cases in which a prediction task is to identify which patients are readmitted within 30 days (i.e., records with “<30” class value are regarded as positive while others negative) or after 30 days (i.e., records with “>30” class value are regarded as positive while others negative), respectively. In these figures, a green line represents the lift chart of ANN while a red and a blue line represents the lift chart of an ideal and a random model, respectively.

According to Figure 1-a, when top 10% of samples are chosen for a prediction task, ANN identifies about 35% of positive samples while an ideal and a random model identifies 100% and 10% of positive samples, respectively. Therefore, the lift (or improvement) of ANN over a random model is calculated as 35%/10% = 3.5. According to Figure 1-a, for ANN to identify 80% of patients who are readmitted within 30 days, it will require 42% of samples.

Figure 2: Lift Chart for Male (top) & Female

Two prediction models (ANN as a green line and DT as a purple line) along with an ideal (a red line) and a random model (a blue line) are shown together for a reference purpose. First of all, in terms of predictive performance, ANN performs better than DT. Then we contrasted patients in each subgroup against a baseline group, patients with the same gender but who were not readmitted. We find that female patients who take more Acarbose medicine are likely to readmitted within 30 days (with a probability = 50.47%). Female patients associated with up dose of Chlorpropamide, down dose of Pioglitazone, and steady dose of Chlorpropamide are also likely to be readmitted within 30 days. In contrast, from male patient group, we note that patients who are associated with steady dose of Tolazamide is most likely to be readmitted within 30 days with a probability of 75.27%. Male patients who keep steady dose of Troglitazone and Tolbutamide or down dose of Glimepiride are also likely to be readmitted within 30 days.

4.3 Racial Group Analysis

We also divided data sets into five subgroups—Caucasian (72.79%), African American (20.92%),

In contrast, for African American group, the most significant medicine indicator is Chlorpropamide: African American patients who have been administered with the steady amount of Chlorpropamide are far more likely to be readmitted within 30 days. Note that two medicines, Chlorpropamide and Repaglinide, are most prevalent medications administered to treat T2D.

4.4 Age Group Analysis

It is also speculated that patient subgroups based on their age would reveal different readmission patterns. To test this, we divided the data set into three subgroups based on age: [0-50] (18.49%), [51-70] (39.53%) and [71-100] (41.98%). According to Figure 4-a, DT outperformed ANN in a readmission prediction of patients in [0-50] age group over a range of between top 10% and top 35% of patients, but ANN performed better in other ranges. However, ANN was superior to DT across almost all ranges for age groups of [51-70] (Figure 4-b) and [71-100] (similar to Figure 4-b, hence omitted due to the limited space). Across all age groups, the number of times that patients have been admitted before is one of the strongest (but too obvious) indicators for a task of predicting whether or not patients will be readmitted within 30 days.

Figure 4-a: Lift Chart for Age Group of [0-50]

We also consider that health administrators will be interested in knowing an optimal magnitude of such a readmission intervention program in terms of how many T2D patients to be included in the program to make it financially feasible.

To estimate the financial feasibility of such a readmission intervention program, we first set potential target (diabetes) population to 38.4 million (Centers for Disease Control and Prevention, 2024) and fixed cost (i.e., initial set up costs) to $100 million based on the latest statistics. Note that using different values for the fixed cost does not necessarily affect the implications of the feasibility analysis. We set variable cost (individual cost in Analysis Server) to $300 per individual, an estimate close to the price of a smart watch/health tracking device that each patient in the program receive and use to track various health related (e.g., heartbeat and blood pressure) and lifestyle information (e.g., daily diet, exercise log).

Finally, we consider two different revenue (or cost saving) settings per each correctly predicted patient to be readmitted: $3,000 ($412.9 billion / 38.4 million * 25% ≈ $2,688) and $6,000 $412.9 billion / 38.4 million * 50% ≈ $5,376). That is, we assume that the readmission prevention program will result in the cost saving of direct and indirect costs by correctly identifying and monitoring who are most likely to be readmitted before they suffer from severe complications and readmitted for a long period of times.

These revenue settings were derived from assumptions that a new program may reduce the total costs of $412.9 billion among 38.4 million T2D patients ($306.6 billion for direct medical costs and $106.3 billion for indirect costs, Centers for Disease Control and Prevention, 2024) by 25% (Scenario 1) or 50% (Scenario 2). We also consult another study [23] in which interventions with intensive lifestyle changes could reduce the risk of developing T2D by 58%.

With these setting values, we obtained two profit charts presented in Figure 5-a (Scenario 1) and 5-b (Scenario 2). In Scenario 1 where the intervention program is assumed to reduce the cost by 25%, we find that it is profitable if at least top 5% of patients participate for the program and its profit (or cost saving) is maximized when 30% of T2D patients are included in the program.

Therefore, findings in this section are particularly useful because various findings from many prior studies may provide scientific insights or medicine indicators related to repeated hospital admission of T2D patients from the perspective of medical professionals and biologists, but they do not provide financial and managerial insights from the perspective of local and federal health administrators who would initiate a readmission intervention program for T2D patients.   

6. Conclusion

This paper considers a simple intervention program to minimize the readmission of T2D patients and tests its financial feasibility. To this end, various prediction models were first calibrated to identify who are most likely to be readmitted and extract influential factors on readmission rates. Through subgroup analyses of gender, race and age, we found that patients in each subgroup show distinct responses to clinical medicines doses during their hospitalization, which may impact the probability of readmission after discharge. In addition, we also found that this simple T2D intervention program can be financially feasible as long as highly predictive models are used to select T2D patients who are likely to be readmitted and program administrators determine the number of T2D patients targeted for the prevention program.

In future, we will consider a more realistic and comprehensive T2D intervention program and test its economic feasibility with better estimates of costs and cost savings. In particular, T2D intervention program should be considered and tested for various socio-economic, demographic, environmental, and genetic factors. In addition, such intervention program should be updated and monitored dynamically.

Therefore, the proposed readmission intervention program with clinically effective treatment of inpatients may be implemented as an initiative project of a comprehensive T2D prevention program. Then it will incorporate multiple approaches such as pre-arranged follow-up appointments, individualized discharge planning, patient education and communication with the primary care provider.

10. Saxena, R., Elbers, C.C., Guo, Y., et al. (2012). Large-scale gene-centric meta-analysis across 39 studies identifies type 2 diabetes loci. American Journal of Human Genetics, 90(4), 410–425.

11. Keating, B.J. (2015. Advances in risk prediction of type 2 diabetes: Integrating genetic scores with Framingham risk models. Diabetes, 64(5), 1495-1497.

12. Asselbergs, F.W., Guo, Y., van Iperen, E.P., Sivapalaratnam, S., Tragante, V., Lanktree, M.B., ... and Baumert, J. (2012). Large-scale gene-centric meta-analysis across 32 studies identifies multiple lipid loci. American Journal of Human Genetics, 91(5), 823-838.

13. Strack, B., DeShazo, J. P., Gennings, C., Olmo, J. L., Ventura, S., Cios, K. J., and Clore, J. N. (2014). Impact of HbA1c measurement on hospital readmission rates: Analysis of 70,000 clinical database patient records. BioMed Research International. doi:10.1155/2014/781670.

14. Liu, X., Li, Y., Li, L., Zhang, L., Ren, Y., Zhou, H., ... & Wang, C. (2016). Prevalence, awareness, treatment, control of type 2 diabetes mellitus and risk factors in Chinese rural population: The RuralDiab study. Scientific Reports, 6, 31426.

15. Harris, M. I., Klein, R., Cowie, C.C., Rowland, M., & Byrd-Holt, D. D. (1998). Is the risk of diabetic retinopathy greater in non-Hispanic blacks and Mexican Americans than in non-Hispanic whites with type 2 diabetes? A US population study. Diabetes Care, 21(8), 1230-1235.

16. Maty, S.C., Everson-Rose, S.A., Haan, M.N., Raghunathan, T.E., & Kaplan G.A. (2005). Education, income, occupation, and the 34-year incidence (1965-99) of Type 2 diabetes in the Alameda County Study. International Journal of Epidemiology, 34(6). 1274-1281. doi:10.1093/ije/dyi167.

17. Brown, A.F., Gregg, E.W., Stevens, M.R., Karter, A.J., Weinberger, M., Safford, M.M., ... & Beckles, G.L. (2005). Race, Ethnicity, Socioeconomic Position, and Quality of Care for Adults with Diabetes Enrolled in Managed Care the Translating Research into Action for Diabetes (TRIAD) study. Diabetes Care, 28(12), 2864-2870.